The dask.array.coarsen documentation provides little detail about the requirements of its reduction argument:

Looking at the relevant part of the implementation:

It seems to me that the requirements on the reduction function are:

• it should be able to take an array as the first argument, and
• it should be able to take in a tuple of axes as an axis= argument, and
• it should be able to not instantiate the input dask array (?). ie does the reduction function of np. only work because of __array_function__ dispatch? Or am I missing some other magic here?

Obviously I'm not super clear on the last one, or whether I'm missing some, and would love some guidance from people more familiar with this function, after which I'd be happy to make a PR or leave this as a good-first-issue for some dask hackathons. 😊